Wednesday, 26 February 2014

The role of Machine Vision as a Rosetta Stone for Artificial Intelligence.

My life has not followed a straight path. There have been many detours and deviations. Never the less, if I turn my head to one side, turn around twice, squint through one eye, and be very selective about what I pick from my memories, I can piece together a narrative that sort-of makes sense.

During my teenage years, I had (typically for a teenager) overwrought and unrealistic ambitions. Driven by a somewhat mystical view of science and mathematics, I harboured ambitions to be a physicist. I wanted to explain and understand a frequently baffling world, as well as (perhaps) to find a way to escape to a place of abstract elegance: A way to remove myself from the social awkwardness and profound embarrassment that plagued my life at that time. I was, after all, a socially inept and acne-ridden nerd with sometimes erratic emotional self-control.

It was in this context that the ethical supremacy of abstract (and disciplined) reasoning over unreliable and sometimes destructive emotional intuition was founded: A concept that forms one of the prime narrative threads that bind this story together.

To me, the abstract world was the one thing that held any hope of making consistent sense; and provided (now as then) the ultimate avenue for a near-perpetual state of denial. Not that I have been terribly successful in my quest (by the overwrought standards of my teenage ambitions at least), but the role of science & technology "groupie" seems to have served me and my career reasonably well so far, and has cemented a view of life as a tragedy in which abstract intellectualism serves as a platonic ideal towards which we forever strive, but are cursed never to achieve.

In any case, I quickly came to the conclusion that my intellectual faculties were completely insufficient to grasp the mathematics that my aspirations required. In retrospect this was less a victory of insight than the sort of failure that teaches us that excessive perfectionism, when coupled with a lack of discipline and determination will inevitably lead to self-imposed failure.

So, I drifted for a few years before discovering Artificial Intelligence, reasoning that if I was not bright enough to be a physicist in my own right, I should at least be able to get some assistance in understanding the world from a computer: an understanding that might even extend to the intricacies of my own unreliable brain. Indeed, my own (possibly narcissistic) quest to improve my understanding both of my own nature and that of the wider world is another key thread that runs through this narrative.

A good part of my motivation at the time came from my popular science reading list. Books on Chaos theory and non-linear dynamics had a great impact on me in those years, and from these, and the notions of emergence that they introduced, I felt that we were only beginning to scratch the surface of the potential that general purpose computing machines offered us.

My (eventual) undergraduate education in AI was (initially) a bit of a disappointment. Focusing on "good old fashioned" AI and computational linguistics, the intellectual approach that the majority of the modules took was oriented around theorem proving and rule-based systems: A heady mix of Noam Chomsky and Prolog programming. This classical and logical approach to understanding the world was really an extension of the philosophy of logic to the computer age; a singularly unimaginative act of intellectual inertia that left little room for the messiness, complexity and chaos that characterised my understanding of the world, whilst similarly confirming my view that the tantalising potential of general-purpose computation was inevitably destined to remain untapped. More than this, the presumption that the world could be described and understood in terms of absolutist rules struck me as an essentially arrogant act. However, I was still strongly attracted to the notion of logic as the study of how we "ought" to think, or the study of thought in the abstract; divorced from the messy imperfections of the real world. Bridging this gap, it seemed to me, was an activity of paramount importance, but an exercise that could only realistically begin at one end: the end grounded in messy realities rather than head-in-the-clouds abstraction.

As a result of this, I gravitated strongly towards the machine learning, neural networks and machine vision modules that became available towards the end of my undergraduate education. These captured my attention and my imagination in a way that the pseudo-intellectualism of computational linguistics and formal logic could not.

My interest in neural networks was tempered somewhat by my continuing interest in "hard" science & engineering, and the lingering suspicion that much "soft" (and biologically inspired) computing was a bit of an intellectual cop-out. A view that has been confirmed a couple of times in my career. (Never trust individuals that propose either neural networks or genetic algorithms without first thoroughly exploring the alternatives!).

On the other hand, machine learning and statistical pattern recognition seemed particularly attractive to my 20-something-year-old mind, combining a level of mathematical rigour which appealed to my ego and my sense of aesthetics, and readily available geometric interpretation which appealed to my predilection for visual and spatial reasoning. The fact that it readily acknowledged the inherent complexity and practical uncertainty involved in any realistic "understanding" of the world struck a chord with me also: It appeared to me as a more intellectually honest and humble practitioners' approach than the "high church" of logic and linguistics, and made me re-appraise the A-level statistics that I had shunned a few years earlier. (Honestly, the way that we teach statistics is just horrible, and most introductory statistics textbooks do irreparable damage to an essential and brilliant subject).

The humility and honesty was an important component. Most practitioners that I met in those days talked about pattern recognition being a "dark art", emphasis on exploratory data analysis and intuitive understanding of the dataset. Notably absent was the arrogance and condescension that seems to characterise the subject now that "Big Data" and "Data Science" have become oh-so-trendy; attracting the mayflies and the charlatans by the boatload.

In any case, then as now, statistical pattern recognition is a means to an end: An engineering solution to bridge the gap between the messy realities of an imperfect world, low level learning and data analysis and the platonic world of abstract thought and logic. This view was reinforced by the approach taken by the MIT COG team: reasoning that in order to learn how to behave in the world, the robot needs a body with sensors and effectors, so it can learn how to make sense of the world in a directed way.

I didn't have a robot, but I could get data. Well, sort of. At that point in time, data-sets were actually quite hard to get hold of. The biggest dataset that I could easily lay my hands on (as an impoverished undergraduate) were the text files from Project Gutenberg; and since my mind (incorrectly) equated natural language with grammars and parsing, rather than statistics and machine learning, my attention turned elsewhere.

That elsewhere was image data. In my mind (influenced by the MIT COG approach), we needed to escape from the self-referential bubble of natural language by pinning abstract concepts to real-world physical observations. Text alone was not enough. Machine Vision would be the rosetta stone that would enable us to unlock it's potential. By teaching a machine to look at the world of objects, it could teach itself to understand the world of men.

One of my fellow students actually had (mirable diu!) a digital camera, which stored its' images on a zip-disk (the size of a 3.25 inch floppy disk), and took pictures that (if I recall correctly) were about 800x600 in resolution. I borrowed this camera and made my first (abortive) attempts to study natural image statistics; an attempt that continued as I entered my final year as an undergraduate, and took on my final year project: tracing bundles of nerves through light microscopy images of serial ultra-microtome sections of drosophila ganglia. As ever, the scope of the project rather outstripped my time and my abilities, but some important lessons were nonetheless learned.

... To be continued.